How Do I Reduce the Ai Score on My Dissertation Without Changing My Argument?
Table of Contents
- Separate Your Argument from Your Prose Surface
- Map Each Chapter's Claims to Evidence
- Revise Wording Without Moving Claims
- Literature Review vs Original Chapters: Different Tactics
- Humanizer on Prose, Not on Your Core Claims
- Preview After Each Chapter Pass
- Argument-Preserving Revision Checklist
- FAQ
- Sources
- Related articles
Separate Your Argument from Your Prose Surface
Your argument is the logical spine: what you claim, what evidence supports it, what you concede, and what you contribute. Your prose surface is how that spine reads on the page—sentence openings, transition phrases, nominalizations, passive vs active voice, and paragraph rhythm.
AI writing detection on student work estimates how closely qualifying prose matches patterns common in large language model output (Turnitin Guides). It does not read your research design, judge your p-values, or know whether Chapter 4’s conclusion is novel. That separation matters because many dissertation revisions that preserve argument still change surface statistics enough to shift the report.
Think of three argument elements you should not “humanize away”:
| Layer | What stays stable | What you may revise on the surface |
|---|---|---|
| Claims | “X moderates Y under condition Z” | Wording of the claim sentence, not the claim itself |
| Evidence | Dataset, instrument, citation keys, numbers | Introductory framing around quoted or paraphrased passages |
| Contribution | What is new relative to the field | How you signpost novelty without hype adjectives |
A common beginner mistake is editing the easiest paragraphs—abstract, limitations, future work—while leaving methods and discussion in the same stiff template voice that triggered flags. Those sections often carry real claims; they deserve surface revision too, but with a claim lock: write the claim in one sentence, revise everything else, then compare before/after to confirm the claim text is unchanged.
Voice drift is another surface problem. Chapter 1 drafted in year one may sound formal and cautious; Chapter 5 drafted after results may sound promotional or chatty. Committees notice fracture; detectors often flag the outlier chapters. Argument-preserving work includes aligning surface rhythm across chapters without merging distinct disciplinary voices—your methods chapter can stay passive-heavy while your discussion stops repeating the same four transition phrases (“Furthermore,” “Moreover,” “In conclusion,” “It is important to note”).
Practical test: read only your topic sentences across the dissertation. If someone could reconstruct your argument from those sentences alone, your spine is intact. Surface revision should not require rewriting topic sentences unless your supervisor already asked for structural change.
Map Each Chapter's Claims to Evidence
Before you touch wording for AI scores, build a claim–evidence map per chapter. This is your insurance policy against accidental argument drift.
For each chapter, list:
- Primary claim (one sentence).
- Sub-claims (bullet list, usually 3–8).
- Evidence type for each sub-claim (your data, a cited study, a definition from the field, a procedural step).
- Risk flag if the prose around that evidence is mostly paraphrase, boilerplate, or tool-smoothed transitions.
Example skeleton for a mixed-methods education dissertation:
- Chapter 2 (Literature review)
- Claim: Gap G exists because prior work treated context C narrowly.
- Evidence: Thematic synthesis of Studies A–F; no new data.
-
Risk: Long paraphrase chains with uniform sentence length.
-
Chapter 3 (Methods)
- Claim: Design D is appropriate for population P and ethics approval E.
- Evidence: Protocol steps, sampling frame, instrument validity citations.
-
Risk: Template protocol language repeated from departmental guides.
-
Chapter 4 (Results)
- Claim: Finding F supports hypothesis H partially.
- Evidence: Tables, figures, statistical tests or coded themes.
-
Risk: Over-explaining numbers in generic academic filler.
-
Chapter 5 (Discussion)
- Claim: Contribution C reframes debate D for practitioners.
- Evidence: Links back to Chapter 2 gap; compares to Studies A–F.
- Risk: Conclusion paragraphs that sound like generic “AI voice” summaries.
Store this map in a spreadsheet or a one-page appendix for your own use—not necessarily for submission. When you revise a flagged paragraph, locate it on the map. If you cannot point to which sub-claim the paragraph serves, the paragraph may be padding; trimming or rewriting it usually lowers AI surface signals and strengthens the thesis.
Supervisors respond well to this map in meetings: “These twelve sub-claims are unchanged; I revised surface wording in sections 3.2 and 4.1 where Turnitin highlighted qualifying prose.” That framing keeps the conversation on scholarship, not on chasing a percentage.
Chapter-level uploads vs merged file: Many students check chapters separately, then merge for the final deposit. Scores can differ when qualifying prose pools across the whole manuscript. Mapping claims per chapter still helps, but add a merge row on your map: which claims must read consistently across the introduction, discussion, and abstract after merge.
Revise Wording Without Moving Claims
Argument-preserving revision uses controlled edits—small enough to audit, large enough to change detectable surface patterns.
The claim-lock paragraph method
- Copy the flagged paragraph into a side document.
- Highlight immutable strings: numbers, named constructs, citation keys, direct quotes, ethics IDs, dataset labels.
- Rewrite connectors and cadence around those strings.
- Diff the before/after: immutable strings should match exactly; everything else may change.
Replace template scaffolding, not substance
Dissertation prose often accumulates empty scaffolding: “This section will discuss…,” “It is widely acknowledged that…,” “The following subsection aims to explore….” Those phrases add length without evidentiary weight and often match high-AI-probability phrasing. Delete scaffolding when the section heading already signals structure; replace with a direct claim sentence tied to evidence.
Vary syntax without changing modality
Keep hedges and strengths stable. If the original says “may suggest,” do not upgrade to “proves.” If the original says “non-significant trend,” do not rewrite as “effect.” Modality is part of your argument; syntax is surface.
Work from evidence outward
For results and discussion paragraphs, draft in this order:
- State the evidentiary fact (statistic, theme quote, observation).
- State the inferential step your method licenses.
- State the boundary (what you do not claim).
Students who draft in reverse—big claim first, evidence later—often produce smooth but hollow prose detectors flag. Evidence-first ordering keeps the argument visible and usually diversifies sentence structure.
Document revisions for integrity conversations
If your institution asks what changed after an AI report, a revision log beats a vague “I rewrote it.” Log: date, chapter, subsection, tool use (if any), and “claims unchanged per map row X.” That log is for you and your supervisor; it is not an admission of misconduct by itself.
Literature Review vs Original Chapters: Different Tactics
Not every chapter plays the same rhetorical role, so one revision tactic for the whole dissertation often moves claims when you only meant to polish prose.
Literature review: synthesis, not novelty theater
In the literature review, your argument is synthetic: you group studies, show tension, and justify your gap. Tactics that preserve argument:
- Anchor every paragraph to at least one named study or clear school of thought.
- Break paraphrase chains: alternate between tight summary, your evaluative sentence (“However, A’s sample limits…”), and a short direct quote where the author’s wording matters.
- Reduce uniform transitions; move logical relations into the subject (“Smith’s RCT, unlike Jones’s survey, …”).
- Avoid faux-originality sentences that claim breakthrough insight in Chapter 2—save contribution language for the discussion.
Do not “humanize” literature into vague opinion (“Many scholars believe learning is important”). That lowers evidentiary precision and can anger supervisors more than a moderate AI flag.
Original chapters (methods, results, ethics-heavy narrative): precision over flourish
Methods and results should sound boring in the right way: repeatable, specific, and tied to protocol. Preserve:
- Step order, inclusion/exclusion criteria, instrument names, software versions.
- All numbers, tables, and figure references.
Surface tactics here:
- Split long passive sentences into two precise sentences.
- Move templates (“Data were analyzed using…”) next to why that test fits your design—one clause of rationale, not a paragraph of filler.
- In qualitative results, keep participant identifiers and code labels exact; revise surrounding interpretive prose.
Introduction and conclusion: align claims, lighten slogans
Introductions and conclusions repeat claims from inner chapters. They are high-risk for repetitive AI-like summaries because students restate the same contribution five ways. Tactic: paste your claim–evidence map’s contribution row, write one contribution paragraph, then delete duplicate summary sentences elsewhere in those chapters.
Literature-heavy chapters may still show similarity flags from legitimate citations; AI flags from uniform smoothing are a different fix. Treat them separately in your revision plan.
If you want to see how chapter-level patterns differ on your manuscript before you merge everything, preview Turnitin reports on the chapter file you are actively revising.
Preview your Turnitin reports before you submit →
Humanizer on Prose, Not on Your Core Claims
An AI humanizer rewrites surface wording while aiming to preserve meaning—useful when you already drafted the argument yourself but transitions and framing sound statistically “generated.” It is a poor tool for inventing claims, filling literature gaps, or rewriting results you have not analyzed.
Safe humanizer zones on a dissertation (when policy allows editing assistance):
- Transition paragraphs between lit-review themes (after citations are locked).
- General introduction and limitations prose after claims are frozen in your map.
- Ethics and procedural narrative that must stay factually identical—humanize only non-numeric sentences and re-verify numbers afterward.
Do-not-humanize zones unless your supervisor explicitly approves:
- Abstract contribution sentences, hypothesis statements, and research questions.
- Statistical results paragraphs, table captions, and quoted participant data.
- Any sentence containing citations you have not checked post-rewrite.
Workflow that protects argument:
- Export or copy only the flagged subsection.
- Run humanizer on that slice.
- Run claim-lock diff against your map.
- Re-read for modality drift (“may” → “will”) and missing citations.
- Paste back into
.docxand confirm formatting if you use a file-based tool.
Humanizers can flatten discipline-specific terms. If your field uses a precise term (“situated cognition,” “intention-to-treat”), add a do-not-change list and manually restore terms after automation.
Never humanize the entire dissertation in one pass: long inputs increase the chance of dropped hedges, merged citations, or broken heading styles. Chapter-sized passes with preview checks beat one bulk rewrite the night before deposit.
Preview After Each Chapter Pass
Argument-preserving revision is iterative, but generic “keep running it until the number drops” without chapter discipline wastes time and risks merge surprises. Replace that with gated previews:
| Gate | When | What you verify |
|---|---|---|
| Gate A | After lit review surface pass | Claims in map still match topic sentences; citations intact |
| Gate B | After methods/results pass | Numbers, labels, and ethics references unchanged |
| Gate C | After discussion pass | Contribution wording consistent with introduction |
| Gate D | After merge to final .docx |
Full-file AI and similarity on the upload you will deposit |
Upload the same file type and structure you plan to submit. If the deposit is PDF generated from Word, preview from that pipeline once before the final week—headings, footnotes, and excluded blocks can shift what counts as qualifying prose.
When a chapter preview drops flags but the merged file spikes, suspect repeated boilerplate across chapters (same limitations paragraph copied, identical transition templates). Fix repetition at the map level: one canonical limitations subsection, cross-reference elsewhere.
If highlights cluster in one theme (e.g., all “significance of the study” paragraphs), batch-revise that theme across chapters so voice aligns—still without changing the underlying claims in your map.
Supervisor alignment preview: send one page—your claim–evidence map plus two before/after paragraphs with highlights—not the whole Turnitin dashboard. Ask: “Are these sub-claims still accurate?” before you spend another week on surface edits.
Argument-Preserving Revision Checklist
Use this checklist as your final pass before deposit. It is designed to lower AI surface signals without moving your thesis spine.
- Build the claim–evidence map for every chapter; mark high-risk sub-claims (long paraphrase, template methods, repetitive conclusions).
- Separate argument from surface in flagged sections; run claim-lock diffs on each rewritten block.
- Apply chapter-specific tactics—synthesis anchors in the literature review, precision edits in methods/results, single contribution paragraph in intro/discussion.
- Delete empty scaffolding and repeated summary sentences that do not add evidence.
- Humanize only permitted prose slices; keep hypotheses, statistics, quotes, and contribution claims out of bulk automation.
- Align voice across chapters without homogenizing disciplinary requirements (methods can stay technical; discussion stops cliché transitions).
- Run gated previews after each major chapter pass and again on the merged deposit file; log changes for supervisor questions.
- Confirm handbook compliance—disclose tools you used, cite AI assistance if required, and keep raw data and analysis files separate from prose polishing.
- Prepare a one-page revision summary for meetings: unchanged claims, edited subsections, no change to evidence tables.
- Read topic sentences aloud; if the argument is clear, surface edits likely preserved what matters.
Before you upload
Step 7 is where merged dissertations surprise students: chapter-level calm scores can combine into a different full-file pattern once the same transition templates repeat across chapters. Preview both similarity and AI on the file you plan to deposit while your map and citations are still editable.
Check your draft for similarity and AI detection →
FAQ
Will rewriting to lower AI risk make my supervisor think I changed my findings?
Not if you document claim-locked edits. Show your claim–evidence map and before/after paragraphs where numbers, quotes, and hypothesis wording are identical. Supervisors worry when results or citations shift—not when you remove template filler or vary transitions.
Should I revise the literature review or results first?
Usually literature review and framing chapters first, because they carry long paraphrase chains and repetitive transitions without touching core findings. Lock results chapters with numeric claim-lock diffs; only then merge and run a full-file preview.
Can I use ChatGPT to “keep my argument” but sound more human?
Institutional policies vary. Where tools are allowed for editing, paste your own draft, instruct the model to preserve claims and citations, and still run claim-lock diffs—models drift modality and drop references. Where tools are forbidden on submitted prose, use manual revision or approved editing support only.
Does lowering AI on one chapter guarantee a lower merged score?
No. Merging can pool qualifying prose and expose repeated boilerplate. Chapter passes plus a merged Gate D preview are both necessary.
Where can I preview reports before my university submission?
You can upload a .docx, .pdf, or .txt draft to a checking service and receive similarity and AI detection Turnitin reports comparable to what many instructors see, typically within minutes. Use previews to guide revision—not as a substitute for supervisor approval.
Sources
- Turnitin Guides — AI writing detection model — qualifying prose, display bands, and false-positive context.
- QAA / institutional AI guidance — consult your own university handbook for permitted editing and disclosure (varies by program).
Bottom line: Reducing AI indicators on a dissertation without changing your argument means locking claims and evidence, revising prose surface with chapter-appropriate tactics, using humanizers only on permitted slices, and previewing the merged file you will actually deposit. The number on the report is a drafting signal; your map and supervisor alignment protect the thesis you must defend.